Models that reason in the native grammar of central banking — reserves, settlement finality, and the issuance of sovereign money. Trained in-nation so the institution that governs the currency also governs the weights.
Fluency is grounded in the actual mechanics of central-bank money, not generic financial text.
The model reasons over reserve accounts, real-time gross settlement, and the distinction between commercial-bank money and central-bank money. It treats settlement finality as a hard concept, not a synonym for payment.
Trained on the lifecycle of sovereign money — issuance against reserves, redemption, and the accounting identities that must hold. It understands that a CBDC liability sits on the central bank's balance sheet, not a vendor's.
The model distinguishes custody from settlement and key-holding from balance-holding. It reasons about a design where the owner holds the keys and the ledger records claims, never conflating the two.
Policy rates, corridors, standing facilities, and open-market operations are first-class concepts. The model can trace how an operational decision propagates to reserve balances and settlement liquidity.
The model is wired to the tamper-evident record so its reasoning is checkable against state.
The model reads from the hash-chained, tamper-evident ledger as ground truth. Claims about balances, transfers, or settlement status resolve to entries an auditor can independently verify.
It reasons natively about delivery-versus-payment as an atomic, all-or-nothing event. The model will not describe a half-settled trade as complete, because the underlying design forbids that state.
Outputs are tied to identifiable ledger entries and rule sources rather than free-form recall. When the model cannot ground a claim, it says so instead of fabricating a figure.
For reconciliation and settlement questions, the model defers to the deterministic engine and explains the result. Language is a layer over the ledger, never a substitute for it.
Ownership and cryptographic assurance extend to the model itself, not only the currency.
Foundation and specialized weights are trained on infrastructure inside the nation's borders, under its data-residency rules. The corpus of monetary and settlement data never leaves the jurisdiction that owns it.
The institution holds the model weights the same way it holds its signing keys — as an asset it controls outright. There is no external tenancy that could revoke or inspect the model.
Model artifacts and the data lineage behind them are signed under ML-DSA-65 (FIPS 204). Integrity of the weights is verifiable against a post-quantum signature, not a vendor's assurance.
Training data provenance, versioning, and evaluation runs are recorded so the model's history is reconstructable. An external auditor can trace what the model was trained on and when.
The model supports the humans who run the currency, without ever moving money on its own.
It drafts policy notes, settlement procedures, and operational memos in the institution's own register. The output is a starting document for expert review, grounded in the current rulebook.
Desk analysts query settlement state, reserve positions, and historical operations in natural language. The model translates the question into ledger queries and returns a cited answer.
The model never initiates issuance, settlement, or transfers. It informs decisions that authorized humans and the deterministic engine execute under key control.
It walks through the mechanical consequences of a proposed operation — how balances, liquidity, and finality would change. The reasoning is transparent so operators can challenge each step.
Talk to us about monetary and sovereign finance in a sovereign deployment.